基于迁移学习的番茄叶片病害图像分类

针对卷积神经网络对番茄病害识别需训练参数较多,训练非常耗时的问题,将迁移学习应用于AlexNet卷积神经网络,对病害叶片和健康叶片共10种类别的番茄叶片进行分类研究。使用14 529张番茄叶片病害图像,随机选择70%作为训练集,30%作为验证集,对AlexNet卷积神经网络模型结构进行迁移,利用在Imagenet图像数据集上训练成熟的AlexNet模型和其参数对番茄叶片病害识别。在训练过程中,固定低层网络参数不变,微调高层网络参数,将番茄病害图像输入到网络中训练网络高层参数,用训练好的模型对10种类别的番茄叶片分类,并进行了20组试验。结果表明:该算法在训练迭代474次时使网络模型很好的收敛,网络对验证集的测试平均准确率达到95.62%,与从零开始训练的AlexNet卷积神经网络相比,本研究算法缩短了训练时间,平均准确率提高了5.6%。采用迁移学习所建立的病害分类模型能够对10种类别的番茄叶片病害快速准确地分类。     

基于Android手机的树木叶片识别系统

为了准确高效地识别树木叶片,开发了一款基于Android 操作平台的树木叶片设别系统。该系统提取13 种树木叶片特征描述子,选择支持向量机作为分类器。该系统包括图像获取、图像处理、特征提取、分类识别和结果展示5 个模块。选取来自15 个树种的1 500 片树叶进行了试验,结果表明,该系统的平均识别率可以达到94.44%,优于BP 神经网络的91.56%,达到了令人满意的效果。该系统特征描述子的筛选、提取以及分类器算法还可以进一步优化,以更好地满足用户需求。     

基于Android手机的树木叶片识别系统

为了准确高效地识别树木叶片,开发了一款基于Android操作平台的树木叶片设别系统.该系统提取13种树木叶片特征描述子,选择支持向量机作为分类器.该系统包括图像获取、图像处理、特征提取、分类识别和结果展示5个模块.选取来自15个树种的1 500片树叶进行了试验,结果表明,该系统的平均识别率可以达到94.44％,优于BP神经网络的91.56％,达到了令人满意的效果.该系统特征描述子的筛选、提取以及分类器算法还可以进一步优化,以更好地满足用户需求.     

基于太赫兹成像技术的大豆叶片水分含量测定

为了实现叶片水分含量的快速、精准检测,提出一种基于太赫兹成像技术的大豆叶片水分含量测定方法。利用太赫兹光谱成像系统获取96份大豆叶片太赫兹图像,采用干燥法测量叶片含水率,通过主成分分析(PCA)提取出水分敏感特征波段0.557、1.098、1.163 THz,对这3个特征波段下的叶片图像采用自适应阈值分割法,将其分为叶脉图像与叶肉图像,分别求取各自的图像灰度特征,并分为叶片特征组(G1)、叶脉特征组(G2)和叶肉特征组(G3)。分别采用多元线性回归(MLR)、反向传播神经网络(BP-ANN)和最小二乘支持向量机(LS-SVM)算法,以上述3个特征组作为输入,构建出9种大豆叶片水分预测模型。对比分析各模型性能,发现基于G3的LS-SVM模型预测结果最好,校正集和预测集的决定系数和均方根误差分别为0.967 8、0.963 2,0.057 8、0.046 5。试验结果表明,利用太赫兹成像技术来检测叶片中的水分含量具有非常高的预测精度,为作物叶片水分含量测定提供了一种行之有效的检测手段。     

基于显著性目标检测的葡萄叶片病害分割

为提高葡萄叶片病害图像中病斑分割性能,提出了一种基于显著性目标检测的病斑分割方法。采用显著性目标检测网络来生成葡萄病害叶片图像的显著性图,通过多种分辨率的网格结构提取图像局部和全局信息,并将它们融合成预测特征;再对病害叶片的显著性图用自适应阈值法分割出叶片上的病害区域,并用形态学方法进行后处理。结果表明,在测试集A上,所建立的方法对病斑分割性能指标马修斯相关系数(MCC)为0.625,略低于对比算法全卷积神经网络(FCN)的0.689,但在衡量泛化性能的测试集B上,所建立方法的MCC为0.338,远高于FCN的0.072, 说明所建立方法在分割精度和泛化性方面具有较好的平衡性。     

Embedded portable device for herb leaves recognition using image processing techniques and neural network algorithm

Herbs have been widely used in food preparation, medicine and cosmetic industry. Knowing which herbs to be used would be very critical in these applications. Nevertheless, the current way of identification and determination of the types of herbs is still being done manually and prone to human error. Designing a convenient and automatic recognition system of herbs species is essential since this will improve herb species classification efficiency. This research focus on recognition approach to the shape and texture features of the herbs leaves. It aims to realize the computerized method to classify the herbs plants in a very convenient way. Portable herb leaves recognition system through image and data processing techniques is implemented as automated herb plant classification system. It is very easy to use and inexpensive system designed especially for helping scientist in agricultural field. The proposed system employs neural networks algorithm and image processing techniques to perform recognition on twenty species of herbs. One hundred samples for each species went through the system and the recognition accuracy was at 98.9%. Most importantly the system is capable of identifying the herbs leaves species even though they are dried, wet, torn or deformed. The efficiency and effectiveness of the proposed method in recognizing and classifying the different herbs species is demonstrated by experiments.     

结合深度学习和引导滤波的苹果叶片图像分割

针对传统方法对苹果叶片进行图像分割和测量几何形状参数精确度较低的问题，结合基于深度学习和引导滤波技术提出一种新的苹果叶片图像自动分割算法。首先采用深度学习方法，使用BiseNet卷积神经网络对苹果叶片图像进行自动分割，得到苹果叶片主体轮廓；然后使用彩色苹果叶片图像作为引导图像对主体轮廓进行引导滤波处理，以增强边缘锯齿等细节特征信息；最后将主体轮廓与细节特征信息进行联合分割，得到完整、准确的苹果叶片信息。对包含174种8 184张苹果叶片图像数据集进行试验，结果表明苹果叶片分割的精确率达到98.99%，交并比98.82%。利用本研究算法能够真正实现准确、快速测量苹果叶片的面积、周长等参数值，为苹果叶片几何参数的测定提供了一种新的测量方法。     

Identification of insect-damaged wheat kernels using short-wave near-infrared hyperspectral and digital colour imaging

Healthy wheat kernels and wheat kernels damaged by the feeding of the insects: rice weevil (Sitophilus oryzae), lesser grain borer (Rhyzopertha dominica), rusty grain beetle (Cryptolestes ferrugineus), and red flour beetle (Tribolium castaneum) were scanned using a near-infrared (NIR) hyperspecrtal imaging system (700-1100 nm wavelength range) and a colour imaging system. Dimensionality of hyperspectral data was reduced and statistical and histogram features were extracted from NIR images of significant wavelengths and given as input to three statistical discriminant classifiers (linear, quadratic, and Mahalanobis) and a back propagation neural network (BPNN) classifier. A total of 230 features (colour, textural, and morphological) were extracted from the colour images and the most contributing features were selected and used as input to the statistical and BPNN classifiers. The quadratic discriminant analysis (QDA) classifier gave the highest accuracy and correctly identified 96.4% healthy and 91.0-100.0% insect-damaged wheat kernels using the top 10 features from 230 colour image features combined with hyperspectral image features.     

基于叶片图像多特征融合的观叶植物种类识别

叶片图像特征提取对于植物自动分类识别有着重要的研究意义。本文以观叶植物叶片为研究对象,综合提取叶片图像的颜色、形状和纹理特征,基于支持向量机(SVM)原理提出了基于图像分析的观叶植物自动识别分类方法。通过对50种观叶植物样本图像进行训练和识别,与BP神经网络和KNN识别方法进行比较,本文所采用的SVM分类器的识别率能够达到91.41%,取得了较好的识别效果。      

基于多特征融合和CNN模型的树种图像识别研究

目的在树种图像识别时会存在类内差异、类间相似的现象,因此导致基于单一人工特征的传统识别方法难以达到理想的识别效果。针对这一问题,本文基于卷积神经网络,提出一种将图像深层特征和人工特征融合的树种图像深度学习识别方法。方法将6类常见树种（樟子松、山杨、白桦、落叶松、雪松和白皮松）图像作为研究对象。首先,通过裁剪、水平翻转、旋转等操作,对原始树种图像集进行数量扩增,并划分为训练集和测试集,建立本次树种识别实验的图像库;其次,将本文模型设计为3路并列网络,分别选取RGB图像、HSV图像、LBP-HOG图像,从图像像素、色彩、纹理和形状的角度出发,对上述树种图像进行识别。一方面构建适合本文实验的CNN深度学习模型,将训练集样本中RGB图像和相对应的HSV图像作为第1路和第2路CNN模型的输入,进行树种图像深层特征提取;另一方面,对训练集进行高斯滤波去噪和人工提取LBP-HOG特征来代表纹理、形状特征,作为第3路CNN模型的输入。然后,将3路模型各自得到的特征在最后一层全连接层进行汇总,作为softmax分类器的最终分类依据。最后,为检验本文方法的可行性,利用上述特征和训练集对SVM分类器、BP神经网络以及现有的深度学习LeNet-5模型、VGG-16模型进行训练,对测试集进行识别验证,来比较最终的识别效果。结果本文提出的多特征融合CNN模型,训练准确率为96.13%,平均验证识别准确率为91.70%。基于单路训练的CNN树种识别模型中,RGB图像作为训练输入值时,识别率最高,为75.21%,HSV特征识别率次之,LBP-HOG特征最差;多特征融合情况下,基于RGB + H通道 + LBP条件下,验证识别准确率最高,达到93.50%;RGB + HSV + LBP + HOG组合识别率不增反降,识别率为89.50%。同样的特征或特征组合条件下,SVM、BP神经网络、LeNet-5模型和VGG-16模型所获得的识别率均低于本文模型的识别率。结论基于RGB + H通道 + LBP特征融合条件下,运用3路并列CNN模型,对本文6类树种图像进行识别的识别率最高,克服了在单一特征情况下识别率低的问题,识别效果也非常理想,实现了从大量不同树种图像中自动识别出具体类别。      

A morphological assessment system for ‘show quality’ bovine livestock based on image analysis

Morphological assessment is one important parameter considered in conservation and improvement programs for bovine livestock. This assessment process consists of scoring an animal based on its morphology and is normally carried out by highly qualified staff. These animals are all of agreed ‘show quality’ and hence they are morphologically very similar.This paper presents a system designed to provide an assessment based on a lateral image of the cow. The system consists of two main parts: a feature extraction stage, to reduce the information on the cow in the image to a set of parameters, and a neural network stage to provide a score based on that set of parameters. For the image analysis section, a model of the animal is constructed by means of the point distribution model (PDM) technique. Later, that model is used in the searching process within each image, which is implemented using genetic algorithms (GAs). As a result of this stage, the vector of weights that describe the deviation of the given shape from the mean is obtained. This vector is used in the second stage, where a multilayer perceptron is trained to provide the desired assessment, using the scores given by experts for selected cows.The system has been tested with 138 images corresponding to 44 individuals of a special rustic breed, with very promising results, given that the information contained in only one view of the cow can not be considered complete.     

Model fusion for prediction of apple firmness using hyperspectral scattering image

Hyperspectral scattering image is an advanced technology widely used in non-destructive measurement of fruit quality. To develop a better prediction model for apple firmness, the present study investigates a model fusion method coupled with wavelength selection algorithms. The current paper first discusses two wavelength selection algorithms, namely, uninformative variable elimination (UVE) and supervised affinity propagation (SAP). The selected effective wavelengths are then set as input to the partial least square (PLS) model. Six hundred “Golden Delicious” apples were analyzed. The first 450 apples were used as sample for the calibration model, whereas the remaining 150 were used for the prediction model. Compared with full wavelengths, the number of effective wavelengths based on the UVE and SAP algorithms decreased to 34% and 35%, but the correlation coefficient of prediction (Rp) increased from 0.791 to 0.805 and 0.814, whereas the root mean-square error of prediction (RMSEP) decreased from 6.00 to 5.73 and 5.71 N, respectively. A fusion model was then developed using UVE-PLS and SAP-PLS models coupled with backpropagation neural network. A better prediction accuracy was achieved from the fusion model (Rp = 0.828 and RMSEP = 5.53 N). The model fusion provides an effective modeling method for apple firmness prediction using hyperspectral scattering image technique.     

Surface grading of bamboo strips using multi-scale color texture features in eigenspace

In order to achieve high competitive quality of bamboo products, it appears that bamboo strips with naturally different tonalities should be elaborately sorted into different classes according to their global color texture appearance. Inspired by the coarse-to-fine visual perception process of human vision system, this paper proposes a new surface grading approach by integrating the color and texture of bamboo strips based on Gaussian multi-scale space. The multi-scale representations of color texture for the original image of bamboo strips could be obtained and used to construct the multivariate image, each channel of which represents a perceptual observation from different scales. The multivariate image analysis (MIA) techniques are used to extract multi-scale features from the resulting multivariate image data. The characteristic images corresponding to typical classes are selected to build the model of the reference eigenspace. The novel testing images and the training images are all projected onto this reference eigenspace to obtain their representative feature clusters. And the Bhattacharyya distance is used to estimate the similarity of the representative feature clusters between the testing images and the training images in the eigenspace. Then a k-NN classifier is adopted to classify the testing images into the given classes of training images. Comparative experiments have been carried out on a set of actual bamboo strip images and the experimental results verify the effective discrimination of multi-scale color texture eigenspace features and good classification accuracy of the proposed surface grading method.     

基于深度学习的5种树皮纹理图像识别研究

目的针对在树皮图像识别时,现有的算法和识别过程过于复杂的问题,提出了基于深度学习的方法来对不同树种的树皮图像进行识别。方法本文以5种常见树种的树皮纹理图像为例,采用基于卷积神经网络的深度学习方法,将原始图像直接作为输入,通过卷积和池化层对图像的低级、高级特征进行自动提取,解决了手动提取纹理特征的困难和问题;在此基础上,对CNN模型结构进行改进,采用带Maxout的ELU激励函数来代替ReLU函数,解决模型的偏移和零梯度问题;对损失函数进行改进,通过添加规范项来优化结构参数,并使用分段常数衰减法对学习率进行动态调控;最后采用softmax分类器对图像类别进行输出。结果对5个树种的树皮图像共计10 000张图像进行实验,其中每类选取200张图像作为测试集。最终训练准确率达到93.80%,测试集识别准确率为97.70%。另外,为验证本文方法的可行性,与传统人工特征提取法,提取HOG特征、Gabor特征和灰度共生矩阵统计法,训练SVM分类器。通过实验比较,本文方法识别准确率最高。结论本文提出的基于深度学习的树皮纹理图像识别方法是可行的,提高了识别效率和精度,为树种的智能化识别提供新的参考。      

大麻对瓜列当和向日葵列当种子萌发诱导作用研究

以宁夏回族自治区固原市当地种植的大麻为材料,采用盆栽试验探索大麻在不同生育期(苗期、快速生长期、开花期)的根际土、根、茎及叶的甲醇和水提取液对瓜列当和向日葵列当种子萌发的刺激效果。结果表明:根际土提取液刺激瓜列当种子的发芽率大于向日葵列当。大麻植株的甲醇和水提取液刺激瓜列当种子发芽率高低顺序为根茎叶,其中甲醇提取液刺激列当种子萌发率高于水提取液。根与茎的甲醇提取液刺激瓜列当和向日葵列当种子发芽率均显著相关(R2=0.833 6,P0.001和R2=0.544 4,P0.05)。植株的甲醇提取液对瓜列当种子的萌发诱导作用在快速生长期最强(45.2%),而向日葵列当种子的萌发诱导作用则在苗期最强(41.5%),水提取液对瓜列当及向日葵列当种子萌发诱导作用均在苗期表现为最强,发芽率分别为53.6%和23.7%。本研究结果表明,大麻在苗期和快速生长期可以作为列当的"捕获"作物,结论可为生物防除寄生杂草列当提供科学依据。     

基于并联卷积神经网络的水果品种识别

为了解决传统的水果图像识别算法在特征提取上的缺陷,以及传统卷积神经网络识别率低的问题,设计了一种基于并联卷积神经网络来提取水果特征的识别方法,利用ELU激活函数替代ReLU激活函数,利用最大类间距损失函数结合传统SoftmaxWithLoss损失函数来提高对相似品种的识别准确率。选取Fruit-360数据集中的8个品种,利用边界均衡生成对抗网络(BEGAN)结合传统的数据增强方法生成大量高质量的数据集,并用其进行训练。结果表明,该模型对8个品种的平均识别准确率达98.85%,具有良好的识别效果。     

Automatic image segmentation method for cotton leaves with disease under natural environment

In order to improve the image segmentation performance of cotton leaves in natural environment, an automatic segmentation model of diseased leaf with active gradient and local information is proposed. Firstly, a segmented monotone decreasing edge composite function is proposed to accelerate the evolution of the level set curve in the gradient smooth region. Secondly, canny edge detection operator gradient is introduced into the model as the global information. In the process of the evolution of the level set function, the guidance information of the energy function is used to guide the curve evolution according to the local information of the image, and the smooth contour curve is obtained. And the main direction of the evolution of the level set curve is controlled according to the global gradient information, which effectively overcomes the local minima in the process of the evolution of the level set function. Finally, the Heaviside function is introduced into the energy function to smooth the contours of the motion and to increase the penalty function Φ(x) to calibrate the deviation of the level set function so that the level set is smooth and closed. The results showed that the model of cotton leaf edge profile curve could be obtained in the model of cotton leaf covered by bare soil, straw mulching and plastic film mulching, and the ideal edge of the ROI could be realized when the light was not uniform. In the complex background, the model can segment the leaves of the cotton with uneven illumination, shadow and weed background, and it is better to realize the ideal extraction of the edge of the blade. Compared with the Geodesic Active Contour(GAC) algorithm, Chan-Vese(C-V) algorithm and Local Binary Fitting(LBF) algorithm, it is found that the model has the advantages of segmentation accuracy and running time when processing seven kinds of cotton disease leaves images, including uneven lighting, leaf disease spot blur, adhesive diseased leaf, shadow, complex background, unclear diseased leaf edges, and staggered condition. This model can not only conduct image segmentation of cotton leaves under natural conditions, but also provide technical support for the accurate identification and diagnosis of cotton diseases.     

A new approach to aflatoxin detection in chili pepper by machine vision

Aflatoxins are the toxic metabolites of Aspergillus molds, especially by Aspergillus flavus and Aspergillus parasiticus. They have been studied extensively because of being associated with various chronic and acute diseases especially immunosuppression and cancer. Aflatoxin occurrence is influenced by certain environmental conditions such as drought seasons and agronomic practices. Chili pepper may also be contaminated by aflatoxins during harvesting, production and storage. Aflatoxin detection based on chemical methods is fairly accurate. However, they are time consuming, expensive and destructive. We use hyperspectral imaging as an alternative for detection of such contaminants in a rapid and nondestructive manner. In order to classify aflatoxin contaminated chili peppers from uncontaminated ones, a compact machine vision system based on hyperspectral imaging and machine learning is proposed. In this study, both UV and Halogen excitations are used. Energy values of individual spectral bands and also difference images of consecutive spectral bands were utilized as feature vectors. Another set of features were extracted from those features by applying quantization on the histogram of the images. Significant features were selected based on proposed method of hierarchical bottleneck backward elimination (HBBE), Guyon’s SVM-RFE, classical Fisher discrimination power and Principal Component Analysis (PCA). Multi layer perceptrons (MLPs) and linear discriminant analysis (LDA) were used as the classifiers. It was observed that with the proposed features and selection methods, robust and higher classification performance was achieved with fewer numbers of spectral bands enabling the design of simpler machine vision systems.     

基于卷积神经网络的小样本树皮图像识别方法

针对在树皮图像分类过程中图像训练数据数量少、识别准确率低的问题,提出一种基于卷积神经网络的小样本树皮图像识别方法。以5种常见树种的树皮图像作为研究对象,在基于卷积神经网络的Inception_v3模型基础上,对原始数据集进行数据增强的一系列操作,扩大数据集的数量;在此基础上,对所有数据集进行白化处理,以降低数据之间的冗余性,使得特征之间相关性较低;采用ReLU激励函数和Dropout方法,防止训练时引起的过拟合现象;同时,在模型的最后添加3层全连接层,增强模型的特征表达能力,采用softmax分类器。最终确定了一个10层CNN模型:5个卷积层、2个池化层、3个全连接层。结果表明,上述网络模型对数据集的识别准确率为94%,并且为验证本研究方法的可行性,分别在MNIST数据集、ImageNet数据集、CIFAR-10数据集进行测试,识别准确率分别为92%、90%、93%。因此,提出的方法在小样本的识别试验中具有较高的识别准确率和一定的可行性。     

A novel algorithm for damage recognition on pest-infested oilseed rape leaves

Cabbage caterpillar infestation of oilseed rape will leave wormholes on leaves. The percentage of wormholes’ area on leaf is an effective index to evaluate infestation seriousness. Hyperspectral imaging technology can be used to extract leaf from non-vegetation objects efficiently. Wormhole reconstruction can then be carried out for counting the wormholes’ area. The reconstruction of wormholes that are entirely within the leaf contour can be easily processed by holes filling function. However, it is difficult to process wormholes at the edge of a leaf. A novel location factor and an improved genetic-wavelet neural network reconstruction algorithm (G-WNNRA) have been proposed in this paper to process wormholes at the edge of a leaf. For the edge of a damaged leaf, the infested part represented by a hole at the edge and non-infested part should be distinguished automatically. Thus the novel location factor which was based on the first derivative of inverse function was used to develop test function for locating the infested part. Then the proposed G-WNNRA was constructed to reconstruct the missing part of an edge following the step of learning the non-infested part of the edge. The topological structure and parameters of the G-WNNRA was optimized by genetic algorithm and morlet wavelet function was applied as a transfer function. The points on non-infested part of edge were adopted as the training data set and the missing part of the edge were predicted. During the prediction, the points making up the reconstructed edge were chosen based on the output of the G-WNNRA. For performance comparison, wavelet neural network (WNN), genetic neural network (GNN) and back propagation neural network (BPNN) were tested on infested oilseed rape leaves and the RMSE of G-WNNRA was smaller than those of WNN, GNN and BPNN. The proposed location algorithm and G-WNNRA can be combined to reconstruct infested oilseed rape leaves.